Architecture of Dynamically Expandable Interactive Trajectory Predictor (IMAGE)
Caption
The model consists of multiple experts with lateral connections. For each new task, a new expert is initialized and trained on the current task dataset. Then the new expert is compared with previous experts to determine whether to expand, while the expert selector is updated to choose the appropriate expert during testing. For task-free continual learning settings, familiarity autoencoder is also trained for each task.
Credit
Communications in Transportation Research
Usage Restrictions
News organizations may use or redistribute this image, with proper attribution, as part of news coverage of this paper only.
License
Original content